Active-learn classification prompt rules

mine type signals from grant docs, synthesize classification rules

improving classifiers via data-driven signal mining from real failures

• · args: JSON object or array per workflow
• · BATCHES: file paths with document data

• · novel classification rules, confusion matrix analysis

• · agent orchestration
• · parallel execution

• · return schema defined and enforced
• · phase logging initialized

• · agent null on error/skip/rate-limit

• · strict return schema enforced
• · saturation analysis proves coverage